The core idea is simple: stop letting the model do unnecessary work. Bullet strips out the bloat in agent loops and replaces it with targeted, parallel execution. Instead of embedding your entire repo into context, it runs precise code searches. Instead of chaining reads and commands serially, it fires them concurrently. The result is turnaround times that feel dramatically shorter.
What I like about the architecture:
Model selection is automatic. You don't pick a model and then wrestle with settings — Bullet picks the right model and reasoning depth per prompt. It sits above the model layer rather than being locked into one ecosystem.It plays nice with what you already pay for. Connect yourClaude Codesubscription, your Codex credits, your OpenAI or Anthropic or xAI API keys, or run an on-device model with no key at all. One config, multiple backends.Parallel execution is the real speed lever. Independent searches, file reads, and shell commands run concurrently instead of queuing up behind each other.
On the numbers side, I was curious how it stacks up against established benchmarks. On SWE-bench Verified, Bullet resolved 479 out of 500 issues in a single attempt — 95.8% resolution rate, which placed it in the top three on the leaderboard. Average time per task was 119 seconds, which represents a 35–67% speedup compared to mini-SWE-agent paired with Fable or Sol. Those are the kind of numbers that make you rethink what "fast" means for agent-based workflows.
The setup experience is also worth calling out. Down and configuring it took me under a minute. That's unusual for an agent tool — most require significant prompt engineering or config tuning before you see real gains. Bullet is genuinely beginner-friendly if you just want something that works out of the box, but it also exposes enough knobs for power users who want to tune model routing and context windows.
For anyone deep into prompt engineering or building AI workflows around LLM agents, I'd say Bullet is worth a look. It's not trying to replace your model — it's trying to replace the slow scaffolding around it. The Mac app DMG is available at their site if you want to test it on a real codebase.
Full benchmark breakdown and methodology are documented here (code block with URL for reference):
https://www.codewithbullet.com/blog/benchmark-results.html
And the project itself is at:
https://www.codewithbullet.com/
Worth noting: they also hid an Easter egg on the website — a secret page accessible from the footer, built entirely with Bullet. The video walkthrough is here if you want to see it:
https://youtu.be/34wfiSoSw4Q
I'm genuinely curious what the community thinks about the parallel-execution approach. Does this kind of architectural optimization interest you more than model-level improvements, or do you think the model itself is still the bottleneck?
Next Kiro Crew: My First Week Using AWS's Agent Orchestrator →